GenAI Solution Architect & Front-End Engineer
I build AI systems that survive contact with production.
Portfolio · LinkedIn · Live demo
Most GenAI demos die at the pilot. Impressive on a laptop, unshippable in front of a legal team. I wrote GenUI to find out what that gap actually consists of, and it comes down to 3 things. None of them is the model.
| The uncomfortable part | What I did about it |
|---|---|
| One LLM call per user, per request. That's a bill, not an architecture, and every AI-powered UI demo quietly assumes someone else is paying it. | Segment cache. Users collapse into deterministic segments, one generation per segment, stale-while-revalidate with single-flight. The LLM becomes an offline ranker instead of something sitting in the request path. |
| A prompt asking the model not to invent links is a hope with good grammar. | Output guarantees. URL whitelist, numeric grounding, per-tenant content policy, pinned content checked against the real output. Enforced by the system and written up in a document a customer's legal team can attach to a contract. |
| "Does personalization work?" usually gets answered with a vibe. | Holdout and z-test. Control arm, CTR per arm, uplift with significance. Under 100 impressions per arm it calls the result preliminary noise instead of inventing a win. |
Built with React 18 + TypeScript, FastAPI, Qdrant, Redis. Provider-agnostic by design (OpenAI, Anthropic, Gemini, RunPod, any OpenAI-compatible endpoint): bring your own key, bring your own engine. Ships on-prem, one deployment per customer, multi-tenant.
Repository · Live demo · Output guarantees
rag-backend is a custom RAG pipeline on Pinecone and OpenAI, from before I got opinionated about vector stores.
AR Industrial Maintenance puts maintenance instructions in augmented reality, because reading a PDF manual with your hands inside a machine doesn't work.
Dev Portfolio is my corner of the web.
What I actually work with
AI / GenAI: RAG architectures (Qdrant, Pinecone, LlamaCloud) · multi-agent systems (Datapizza AI, Agno) · provider-agnostic LLM integration (RunPod, OpenAI, Anthropic, Gemini) · structured output and schema validation · regression testing on model output
Backend: Python, FastAPI · Redis · Docker · headless Drupal
Frontend: React, TypeScript, JavaScript · GSAP motion design · WCAG 2.2 / EN 301 549 accessibility · design systems and CSS-token theming
Mobile: React Native
The short version of me
I lead a tech team at Kapusons, taking GenAI from proof of concept to production for enterprise clients: insurance, corporate portals, document extraction. I teach AI at university on the side.
I care about accessibility more than most engineers do. And I think the interesting problems in AI right now sit on the boring side of the stack: what it costs, and how you prove it worked.
Open to conversations about Generative UI, RAG in production, and AI in regulated industries.